URCS Systems Group Seminar: November 16, 2018

Design and Analysis of Large-Scale GPU Architectures

Onur Kayiran, AMD

2:00 – 3:30 PM, Wegmans Hall 2506 conference room

The total market for Graphics Processing Units (GPUs) is growing due to increasing demand for compute power and more capabilities in various areas such as gaming, block-chain, high-performance computing, and machine learning. The increasing demand for higher compute power resulted in a trend of larger die sizes with newer generation of GPUs. Such large-scale GPUs pose several challenges for architects. In this talk, I will focus on two of these challenges.

In the first part of this talk, I will discuss how the simulation of interconnect and the memory system becomes a major challenge in large-scale GPU systems, and then present a simulation methodology for evaluating the memory system for large-scale GPUs. In the second part of this talk, I will focus on the increasing importance of on-chip interconnect and cache bandwidth as well as on-chip data movement in large-scale GPUs. Then, I will discuss how data requested by a Compute Unit (CU) can be obtained from another CU in the presence of inter-CU data locality, and present techniques that unlock this additional potential source of bandwidth by effectively coordinating data movement across CUs.

Students, faculty, and staff from UR and neighboring institutions are welcome to join us for off-campus lunch prior to the talk from noon – 2:00. Lunch is self-paid and typically costs about $12–$15.

The lunch group departs at 12:05 PM from the Wegmans Hall third floor lobby.
Please email the systems@cs.rochester.edu mailing list if you plan to attend so we can plan rides.